rbindlist: unprotect longestLevels inside the loop (#7794)

* test case

* Protect coercedForFactor deeper in the stack

* Unprotect longestLevels inside the loop

* NEWS entry

* Keep listNames protected

* Clarify test comment

Co-Authored-By: Michael Chirico <chiricom@google.com>

* Detect --max-ppsize

Co-Authored-By: Benjamin Schwendinger <52290390+ben-schwen@users.noreply.github.com>

---------

Co-authored-by: Michael Chirico <chiricom@google.com>
Co-authored-by: Benjamin Schwendinger <52290390+ben-schwen@users.noreply.github.com>

Author: 3 people

NEWS.md

@@ -58,7 +58,9 @@
 
 12. `print.data.table()` now truncates long character columns and list-column summaries by default to avoid horizontal console overflow, [#7718](https://github.com/Rdatatable/data.table/issues/7718). When `datatable.prettyprint.char` is `NULL` (the default), the truncation limit is now dynamically calculated based on the available console width. Use `options(datatable.prettyprint.char=Inf)` for the old default behavior (never truncate). Thanks @tdhock for the report and @venom1204 for the fix.
 
-13. `as.IDate()` and `as.ITime()` now preserve names, matching base `as.Date()` behavior, [#7252](https://github.com/Rdatatable/data.table/issues/7252). Thanks @DavisVaughan for the report and @venom1204 for the PR.
+13. `rbindlist()` (and therefore the `rbind()` method for `data.table`s) no longer raises an error upon encountering more than approximately 50000 columns in a list entry, [#7793](https://github.com/Rdatatable/data.table/issues/7793). The bug was introduced in `data.table` version 1.18.2.1. Thanks to @rickhelmus for the report and @aitap for the fix.
+
+14. `as.IDate()` and `as.ITime()` now preserve names, matching base `as.Date()` behavior, [#7252](https://github.com/Rdatatable/data.table/issues/7252). Thanks @DavisVaughan for the report and @venom1204 for the PR.
 
 ### Notes
 

inst/tests/tests.Rraw

@@ -21670,14 +21670,29 @@ test(2375.4, print(data.table(x="abcdefghijklmnopqrstuvwxyz")), output="abcdefgh
 test(2375.5, print(data.table(id=1L, score=99.1, txt="abcdefghijklmnopqrstuvwxyz")), output="abcdefghijklmn...", options=list(width=20, datatable.prettyprint.char=NULL))
 test(2375.6, print(data.table(x=rep("ABCDEFGHIJKLMNOPQRSTUVWXYZ", 1e6)), topn=1), output="1000000: ABCDEFGHIJKLM...", options=list(width=25, datatable.prettyprint.char=NULL))
 
+# rbindlist() used to put O(ncol(...)) > R_PPStackSize items on the protect stack, #7793
+# Assuming 50000 as the default protection stack size, although R --max-ppsize=... can increase it to 500000.
+max_ppsize = local({
+  arg = grep("^--max-ppsize=", commandArgs(FALSE), value = TRUE)
+  if (length(arg)) {
+    v = suppressWarnings(as.integer(sub("^--max-ppsize=", "", tail(arg, 1L))))
+    if (!is.na(v)) return(v)
+  }
+  50000L
+})
+x = as.data.table(as.list(1:max_ppsize))
+test(2376, rbindlist(list(x)), x)
+rm(x, max_ppsize)
+
 # #7252 as.IDate()/as.ITime preserve names
-test(2376.01, names(as.IDate(c(a = "2019-01-01"))), "a")
-test(2376.02, names(c(a = as.IDate("2019-01-01"))), "a")
-test(2376.03, names(as.ITime(c(a = "12:00:00"))), "a")
-test(2376.04, names(as.IDate(structure(as.POSIXct("2019-01-01 12:00:00"), names = "a"))), "a")
-test(2376.05, names(as.ITime(structure(3600, names = "a"))), "a")
-test(2376.06, names(as.IDate(c(a = 18000))), "a")
-test(2376.07, names(as.IDate(c(a = 1), origin = "2020-01-01")), "a")
-test(2376.08, names(as.ITime(c(a = "12-00-00"), format = "%H-%M-%S")), "a")
-test(2376.09, names(as.IDate(as.POSIXct(c(a = "2019-01-01"), tz="UTC"))), "a")
-test(2376.10, names(as.IDate(as.POSIXct(c(a = "2019-01-01"), tz="America/New_York"))), "a")
+test(2377.01, names(as.IDate(c(a = "2019-01-01"))), "a")
+test(2377.02, names(c(a = as.IDate("2019-01-01"))), "a")
+test(2377.03, names(as.ITime(c(a = "12:00:00"))), "a")
+test(2377.04, names(as.IDate(structure(as.POSIXct("2019-01-01 12:00:00"), names = "a"))), "a")
+test(2377.05, names(as.ITime(structure(3600, names = "a"))), "a")
+test(2377.06, names(as.IDate(c(a = 18000))), "a")
+test(2377.07, names(as.IDate(c(a = 1), origin = "2020-01-01")), "a")
+test(2377.08, names(as.ITime(c(a = "12-00-00"), format = "%H-%M-%S")), "a")
+test(2377.09, names(as.IDate(as.POSIXct(c(a = "2019-01-01"), tz="UTC"))), "a")
+test(2377.10, names(as.IDate(as.POSIXct(c(a = "2019-01-01"), tz="America/New_York"))), "a")
+test(2376.11, names(as.IDate(c(a = 1), origin = c(b = "2020-01-01"))), "b")

src/rbindlist.c

@@ -250,7 +250,7 @@ SEXP rbindlist(SEXP l, SEXP usenamesArg, SEXP fillArg, SEXP idcolArg, SEXP ignor
   setAttrib(ans, R_NamesSymbol, ansNames);
   if (idcol) {
     SET_STRING_ELT(ansNames, 0, STRING_ELT(idcolArg, 0));
-    SEXP idval, listNames=getAttrib(l, R_NamesSymbol);
+    SEXP idval, listNames=PROTECT(getAttrib(l, R_NamesSymbol));
     if (length(listNames)) {
       SET_VECTOR_ELT(ans, 0, idval=allocVector(STRSXP, nrow));
       for (int i=0,ansloc=0; i<LENGTH(l); ++i) {
@@ -270,16 +270,19 @@ SEXP rbindlist(SEXP l, SEXP usenamesArg, SEXP fillArg, SEXP idcolArg, SEXP ignor
         for (int k=0; k<thisnrow; ++k) idvald[ansloc++] = i+1;
       }
     }
+    UNPROTECT(1); // listNames
   }
 
-  SEXP coercedForFactor = NULL;
+  PROTECT_INDEX IcoercedForFactor;
+  SEXP coercedForFactor = R_NilValue;
+  PROTECT_WITH_INDEX(coercedForFactor, &IcoercedForFactor); nprotect++;
   for(int j=0; j<ncol; ++j) {
     int maxType=LGLSXP;  // initialize with LGLSXP for test 2002.3 which has col x NULL in both lists to be filled with NA for #1871
     bool factor=false, orderedFactor=false;     // ordered factor is class c("ordered","factor"). isFactor() is true when isOrdered() is true.
     int longestLen=-1, longestW=-1, longestI=-1; // just for ordered factor; longestLen must be initialized as -1 so that rbind zero-length ordered factor could work #4795
     PROTECT_INDEX ILongestLevels;
     SEXP longestLevels=R_NilValue;              // just for ordered factor
-    PROTECT_WITH_INDEX(longestLevels, &ILongestLevels); nprotect++;
+    PROTECT_WITH_INDEX(longestLevels, &ILongestLevels);
     bool int64=false, date=false, posixct=false, itime=false, asis=false;
     const char *foundName=NULL;
     bool anyNotStringOrFactor=false;
@@ -349,7 +352,7 @@ SEXP rbindlist(SEXP l, SEXP usenamesArg, SEXP fillArg, SEXP idcolArg, SEXP ignor
     if (factor && anyNotStringOrFactor) {
       // in future warn, or use list column instead ... warning(_("Column %d contains a factor but not all items for the column are character or factor"), idcol+j+1);
       // some coercing from (likely) integer/numeric to character will be needed. But this coerce can feasibly fail with out-of-memory, so we have to do it up-front
-      if (coercedForFactor==NULL) { coercedForFactor=PROTECT(allocVector(VECSXP, LENGTH(l))); nprotect++; }
+      if (coercedForFactor==R_NilValue) REPROTECT(coercedForFactor=allocVector(VECSXP, LENGTH(l)), IcoercedForFactor);
       for (int i=0; i<LENGTH(l); ++i) {
         SEXP li = VECTOR_ELT(l, i);
         int w = usenames ? colMap[i*ncol + j] : (j<length(li) ? j : -1); // check if j exceeds length for fill=TRUE and usenames=FALSE #5444
@@ -563,7 +566,8 @@ SEXP rbindlist(SEXP l, SEXP usenamesArg, SEXP fillArg, SEXP idcolArg, SEXP ignor
         ansloc += thisnrow;
       }
     }
+    UNPROTECT(1); // longestLevels
   }
-  UNPROTECT(nprotect); // ans, ansNames, longestLevels? coercedForFactor?
+  UNPROTECT(nprotect); // ans, ansNames, coercedForFactor
   return(ans);
 }